Nvidia's flagship AI rack slips to 2028 — while the memory chipmakers who feed it print record profits

Information Technology 3 min 80 sources

Nvidia's next big server system is delayed more than a year on a manufacturing snag. The chipmakers who supply the one scarce part it needs — memory — are having the best year of their lives.

Key takeaways

Nvidia's flagship AI server rack, Kyber, has been delayed more than a year to 2028 because of one hard-to-manufacture circuit board.
The memory chipmakers that supply the scarce parts AI needs are having record quarters — Samsung's profit up eighteen-fold, all three big makers now worth over a trillion dollars.
When demand for a product outruns the supply of one essential part, the money flows to whoever owns that part, not to whoever builds the finished thing.

Two numbers landed this week that only make sense side by side.

On Monday, research firm SemiAnalysis said Nvidia’s next marquee product — a server cabinet called Kyber — has slipped by more than 12 months, to 2028 [4]. Kyber packs 144 of Nvidia’s most powerful chips into one unit so they run as a single giant computer, the horsepower AI companies need to train their largest models [4]. It was meant to debut in 2027 with Nvidia’s next-generation Rubin chips [4]. The holdup is one component: a “PCB midplane,” a specialised multi-layer circuit board that wires the modules together, which SemiAnalysis said “remains challenging from a manufacturability standpoint” [4]. A larger version linking eight racks is likely delayed too [4].

The same week, the companies that make the memory chips those racks depend on reported the best quarters of their lives.

The suppliers are having a record year

Samsung is expected to flag operating profit of about 86 trillion won — roughly $56 billion — for the April-to-June quarter, an eighteen-fold jump from a year earlier [5]. The reason is a memory shortage: AI systems consume enormous amounts of high-bandwidth memory, the fast chips that sit next to the processor and feed it data, and demand has outrun supply, pushing prices up [5]. Shares of the three big memory makers — Samsung, SK Hynix, and America’s Micron — are up 158%, 273%, and 242% this year [5]. All three are now worth more than a trillion dollars [5].

On Monday, SK Hynix launched an American share sale of about $28 billion — one of the world’s largest — selling 17.79 million new shares as depositary receipts to tap US investors directly [3][6]. SK Hynix is a key supplier of high-bandwidth memory to Nvidia and Google [3]. For years it traded at a discount to Micron; the listing is a bid to close that gap by reaching “the world’s deepest equity market and its frenzy for all things AI,” as Bloomberg put it [6]. Last week the company also said it would spend 100 trillion won — about $64 billion — on new chip plants [3].

Foxconn, which assembles much of the AI server hardware, reported second-quarter revenue up 40% from a year earlier [7].

The pattern under the two numbers

The finished product is stuck. The suppliers of its scarcest input are being showered with money. That is not a contradiction — it is the same fact seen from two ends. When demand for a complex product outruns the supply of one essential part, the value flows to whoever owns that part, not to whoever builds the whole thing. Nvidia designs the rack; the memory makers own the choke it runs through, and the market is paying them for it.

South Korea has noticed. On Sunday, a top presidential official said the government will build a “Future Response Fund” from the extra tax revenue the chip boom is throwing off, aimed at national investment projects and at narrowing inequality [9]. When a windfall lands this concentrated, governments start asking who else should get a share.

Also worth knowing

Amazon is winding down Mechanical Turk, the crowdsourcing marketplace it launched in 2005 [10]. From July 30 it stops taking new customers; existing ones can stay, but Amazon says it will add no new features [10]. Mechanical Turk paid people small sums to do micro-tasks — labelling images, checking data — and became a workhorse for annotating the data that trained early AI models [10]. It fell into decline as bots, fraud, and workers quietly using AI to do the tasks eroded trust in the results [10]. The platform that helped humans train the machines is closing as the machines take over the tasks.

A note of caution runs through the whole picture. On Saturday, chip stocks slumped before the US holiday, and even a bull like CNBC’s Jim Cramer conceded “there is not a day when someone tries to rain on Nvidia’s parade” [1]. A boom priced this steeply reacts to every wobble.

02 · Lesson · why it matters

The money doesn't go to whoever finishes the thing — it goes to whoever owns the part that's hard to get

When something can't be built fast enough, the profit pools at the tightest link in the chain, not at the end of it.

Two stories that are one story

Nvidia’s next great server rack is late. The thing that makes AI systems learn — a cabinet of 144 chips working as one — slipped more than a year, held up by a single circuit board nobody can build fast enough. That sounds like bad news for the AI boom.

In the same week, the companies that make the memory chips those racks run on posted the best quarters they have ever had. One expects to have earned eighteen times last year’s profit. Three of them are now worth more than a trillion dollars each.

The finished product is stuck. The suppliers of one of its parts are drowning in money. Read them apart and they look like opposite news. Read them together and they are the same fact told from two ends of a chain.

Where the money pools

Picture the whole thing as a line. Raw materials, then the memory chips, then the processors, then the rack, then the data centre, then the AI service someone finally uses. Demand is pouring in at the far end — everyone wants the AI. But you cannot buy the far end without buying everything upstream of it.

So the demand travels back up the line, looking for something to buy. And it keeps travelling until it hits the narrowest gate — the one part there isn’t enough of. Right now that gate is high-bandwidth memory, the fast chips that feed data to the processor. Everyone building AI needs it, and there isn’t enough. So the price sits there and rises.

The value doesn’t spread evenly along the chain. It pools at the tightest link. Not because the memory makers are cleverest or work hardest, but because they happen to sit on the thing that’s scarce. The company that assembles the finished rack can be brilliant and still capture less, because anyone with the parts can assemble a rack — but only a few firms can make the memory.

This isn’t only about chips

Name the pattern and you start seeing it everywhere. In a gold rush, the people who reliably got rich were the ones selling shovels and denim — the scarce, necessary thing every miner needed — not most of the miners. When a city booms, the builders and the buyers change, but the ground stays put and the landowner collects. In a hospital short of one specialist, the whole schedule bends around that one person’s hours.

The rule holds: find the step that can’t be scaled up quickly, and you’ve found where the reward collects. Effort and cleverness matter, but they lose to position. Owning the bottleneck beats being good at the parts that aren’t scarce.

Who’s inside this

It’s easy to read this as a story about giant companies and skip past your own seat in it. Don’t. If you have a pension or savings in a broad market fund, some of this year’s gains — Samsung up 158%, SK Hynix up 273% — are quietly yours, because those chipmakers now move the whole index. The AI you’ll use next year costs what it costs partly because of who owns the memory today. And the same logic runs through your own work: in any team, any market, any craft, the leverage sits with whoever holds the part that’s hard to replace, and it drains away from the part that anyone can do. That’s worth knowing about your own position, not just Nvidia’s.

The shape someone built

None of this fell from the sky. Memory being the choke point isn’t a law of nature — it’s the result of years of decisions about who built which plants, who was allowed to sell where, which government backed which industry. South Korea spent decades pouring public support into these companies, and this week its government said it will take a slice of the windfall through a new fund, because when a boom lands this concentrated, the question of who else deserves a share arrives fast.

The arrangement serves the chipmakers, plainly. It can also serve the people under it — the jobs, the tax revenue, the fund meant to spread the gains. Both are true at once. The point isn’t that someone is a villain. The point is that the map of who-gets-rich was drawn, not born, and it could be drawn differently.

On the whole

The lesson is small and useful: when a thing can’t be built fast enough, don’t watch the finished product to see where the money goes — watch the tightest link. That’s where it pools.

But hold it loosely. Today’s choke point is memory; a new plant, a new design, a better circuit board, and the narrow gate moves somewhere else, and the money follows it there. Nobody sitting at one link — not the chipmaker, not Nvidia, not the investor, not you — can see the whole chain or how fast it will shift. We each see our own gate clearly and the rest as rumour. Knowing that the money pools at the bottleneck is worth something. Knowing you can’t reliably predict which link is next is worth more.

03 · Lab · your turn

Where The Money Pools

Rehearse betting on a supply chain and watch profit collect at the scarce link, not the finished product.

04 · Hope · carry this

A bottleneck is not a wall — it is a signal, flashing to everyone with money and skill that this is the thing we need more of. That signal is exactly how shortages end: the reward for solving them is what pulls the next plant, the next design, the next fix into being.